Classification of Saffron Using Color Features Extracted from the Image

The classification of saffron as the most expensive spice is of great importance for customers and traders. In general, two methods are currently used to classify saffron. The first method is based on the experiences of an expert and by observing the samples. The second method is destructive and is performed using laboratory methods. According to experts, the use of machine learning techniques to classify saffron is a goal due to its non-destructive nature and timely characteristics. This method can also increase the accuracy of the industrial scale grading process. In this paper, a vision machine method is presented. Due to lack of documented research on this subject, a comprehensive literature search is presented in this work. Almost all color characteristics were extracted and used in a large number of classifiers. Experts in Iran classify saffron into three main categories based on their appearance: Pushal, Negin and Sargol. In this paper, a database consisting of 440 images from saffron for the three different classes was collected using a mobile phone camera. After applying some preprocessing steps, such as background removal, cropping etc., 21 color features were extracted using different image analysis methods. Twenty-two classifiers were employed for classification. Comparing results of different classifiers showed that the Linear Discriminant, Linear SVM, Bagged Trees and RUSBoost Trees can produce more accurate grading compared to other classifiers when using color features. In particular, mean classification accuracy of 82.23% was achieved in this work using Linear a SVM classifier.